Added official file support

src/llm.ts

@@ -5,11 +5,16 @@ import {OpenAi} from './open-ai.ts';
 import {LLMProvider} from './provider.ts';
 import {AiTool, AiToolArg} from './tools.ts';
 import {fileURLToPath} from 'url';
-import {dirname, join} from 'path';
 import {spawn} from 'node:child_process';
 import {Memory, MemoryCache, MemoryManager, MemoryOptions, stripHeader} from './memory.ts';
+import {mkdtempSync} from 'node:fs';
+import fs from 'node:fs/promises';
+import {tmpdir} from 'node:os';
+import {dirname, join, basename, extname} from 'path';
+import { PDFParse } from 'pdf-parse';
 
 const MAX_AGENT_DEPTH = 5;
+const PDF_OCR_PAGE_THRESHOLD = 12; // above this many pages, OCR scanned pages instead of feeding images to the model
 
 export type AnthropicConfig = {proto: 'anthropic', token: string | string[]};
 export type OpenAiConfig = {proto: 'openai', host?: string, token: string | string[]};
@@ -27,6 +32,19 @@ export type Agent = {
 	agents?: string[] | null;
 }
 
+export type LLMFile = {
+	/** Path to file on disk */
+	path?: string;
+	/** File content: raw text, base64-encoded binary, or a Buffer */
+	content?: string | Buffer;
+	/** Original filename, used to infer type from extension */
+	name?: string;
+	/** Mime type override, inferred from extension if omitted */
+	mime?: string;
+	/** @internal set once extraction has run, skips re-processing next turn */
+	extracted?: boolean;
+};
+
 export type LLMMessage = {
 	/** Message originator */
 	role: 'assistant' | 'system' | 'user';
@@ -88,6 +106,8 @@ export type LLMRequest = {
 	mcp?: McpServer[];
 	/** Subagents exposed as delegatable/wrapped tools */
 	agents?: Agent[];
+	/** Attach files to request */
+	files?: LLMFile[];
 	/** @internal recursion guard for nested agent delegation */
 	_agentDepth?: number;
 }
@@ -111,6 +131,11 @@ export type Skill = {
 }
 
 class LLM {
+	private static AUDIO_EXT = ['wav','mp3','m4a','flac','ogg','aac','wma'];
+	private static IMAGE_EXT = ['png','jpg','jpeg','bmp','gif','tiff','webp'];
+	private static TEXT_EXT = ['txt','md','csv','json','xml','html','js','ts','py','yaml','yml','log'];
+	private static PDF_EXT = ['pdf'];
+
 	private memoryManager!: MemoryManager;
 
 	defaultModel!: string;
@@ -126,6 +151,120 @@ class LLM {
 		this.memoryManager = new MemoryManager(this);
 	}
 
+	private async loadBuffer(file: LLMFile, asText: boolean): Promise<Buffer> {
+		if(file.path) return fs.readFile(file.path);
+		if(Buffer.isBuffer(file.content)) return file.content;
+		if(typeof file.content === 'string') return Buffer.from(file.content, asText ? 'utf-8' : 'base64');
+		throw new Error('No path or content provided');
+	}
+
+	private async writeTemp(name: string, buffer: Buffer): Promise<string> {
+		const path = join(mkdtempSync(join(tmpdir(), 'ai-file-')), name);
+		await fs.writeFile(path, buffer);
+		return path;
+	}
+
+	/**
+	 * Extract text from a PDF. Pages with no text layer (scanned/image-only) are handled as either:
+	 * - Rendered to images and returned alongside the text so the (vision-capable) model can read them directly
+	 * - OCR'd via Tesseract when the doc is too large to reasonably pass as images
+	 */
+	private async resolvePdf(buffer: Buffer): Promise<{text: string, images: {mime: string, data: string}[]}> {
+		const parser = new PDFParse({data: buffer});
+		try {
+			const {text, pages} = await parser.getText();
+			const scanned = (pages || []).filter(p => !p.text?.trim());
+			if(!scanned.length) return {text: text.trim() || '[Empty PDF]', images: []};
+			const total = pages.length;
+			const pageNums = scanned.map(p => p.num);
+			const {pages: shots} = await parser.getScreenshot({partial: pageNums});
+			if(total <= PDF_OCR_PAGE_THRESHOLD) {
+				return {
+					text: text.trim(),
+					images: shots.map(s => ({mime: 'image/png', data: Buffer.from(s.data).toString('base64')}))
+				};
+			}
+			const ocrText = await Promise.all(shots.map(async (s, i) => {
+				const path = await this.writeTemp(`page-${pageNums[i]}.png`, Buffer.from(s.data));
+				try {
+					return await this.ai.vision.ocr(path) || '';
+				} finally {
+					fs.rm(dirname(path), {recursive: true, force: true}).catch(() => {});
+				}
+			}));
+			return {text: [text.trim(), ...ocrText].filter(Boolean).join('\n\n'), images: []};
+		} finally {
+			await parser.destroy();
+		}
+	}
+
+	private async resolveFile(file: LLMFile): Promise<{text?: string, images?: {mime: string, data: string}[]}> {
+		const name = file.name || (file.path ? basename(file.path) : 'file');
+
+		// Already resolved on a previous turn, reuse cached text
+		if(file.extracted) return {text: `<file name="${name}">\n${file.content}\n</file>`};
+
+		const ext = extname(name).slice(1).toLowerCase();
+		const mime = file.mime || '';
+		const isAudio = mime.startsWith('audio/') || LLM.AUDIO_EXT.includes(ext);
+		const isImage = mime.startsWith('image/') || LLM.IMAGE_EXT.includes(ext);
+		const isPdf = mime === 'application/pdf' || LLM.PDF_EXT.includes(ext);
+		const isText = mime.startsWith('text/') || LLM.TEXT_EXT.includes(ext);
+
+		let tmpDir: string | null = null;
+		try {
+			if(isImage) {
+				const data = (await this.loadBuffer(file, false)).toString('base64');
+				return {images: [{mime: mime || `image/${ext === 'jpg' ? 'jpeg' : ext}`, data}]};
+			}
+
+			if(isPdf) {
+				const {text, images} = await this.resolvePdf(await this.loadBuffer(file, false));
+				// Only cache/skip re-processing when we didn't need to hand off images (OCR'd or fully text-based)
+				if(!images.length) {
+					file.content = text;
+					file.extracted = true;
+					delete file.path;
+				}
+				return {text: `<file name="${name}">\n${text || '[Scanned PDF - see attached page images]'}\n</file>`, images};
+			}
+
+			let text: string;
+			if(isAudio) {
+				let path = file.path;
+				if(!path) {
+					const buffer = await this.loadBuffer(file, false);
+					path = await this.writeTemp(name, buffer);
+					tmpDir = dirname(path);
+				}
+				text = await this.ai.audio.asr(path) || '';
+			} else if(isText) {
+				text = (await this.loadBuffer(file, true)).toString('utf-8');
+			} else {
+				text = `Unsupported file type: ${ext || mime}`;
+			}
+
+			// Cache result, skip re-extraction on future turns of the same conversation
+			file.content = text;
+			file.extracted = true;
+			delete file.path;
+
+			return {text: `<file name="${name}">\n${text}\n</file>`};
+		} catch(err: any) {
+			return {text: `<file name="${name}">Failed to process: ${err.message}</file>`};
+		} finally {
+			if(tmpDir) fs.rm(tmpDir, {recursive: true, force: true}).catch(() => {});
+		}
+	}
+
+	private async resolveFiles(files: LLMFile[]): Promise<{text: string, images: {mime: string, data: string}[]}> {
+		const resolved = await Promise.all(files.map(f => this.resolveFile(f)));
+		return {
+			text: resolved.filter(r => r.text).map(r => r.text).join('\n\n'),
+			images: resolved.flatMap(r => r.images || [])
+		};
+	}
+
 	private setupAgent(agents: Agent[] = [], allAgents: Agent[], history: LLMMessage[], aborts: (() => void)[], depth = 0, delegateState: {resp: string | null}): AiTool[] {
 		return agents.map(a => {
 			const toolName = `${a.delegate ? '' : 'sub'}agent_${snakeCase(a.name)}`;
@@ -343,6 +482,18 @@ Linked: ${makeUnique([...r.links, ...r.backlinks]).join(', ')}
 
 			if(aborted) throw Object.assign(new Error('Aborted'), {name: 'AbortError'});
 
+			// Files
+			const files = options.files || [];
+			const lastMsg = history[history.length - 1];
+			const originalContent = lastMsg?.content;
+			if(files.length && lastMsg?.role === 'user') {
+				const {text, images} = await this.resolveFiles(files);
+				const merged = text ? `${originalContent}\n\n${text}` : originalContent;
+				lastMsg.content = images.length
+					? [...images.map(i => ({type: 'image', mime: i.mime, data: i.data})), {type: 'text', text: merged}]
+					: merged;
+			}
+
 			const toolTimings = new Map<string, {duration: number, tps: number}>();
 			tools = this.wrapToolTiming(tools, toolTimings);
 
@@ -352,6 +503,8 @@ Linked: ${makeUnique([...r.links, ...r.backlinks]).join(', ')}
 			request = this.models[m].ask('', {...options, tools, system: prompts.filter(Boolean).join('\n\n')});
 			let resp = await request;
 
+			if(files.length && lastMsg?.role === 'user') lastMsg.content = originalContent;
+
 			// Capture meta (duration / tps)
 			for(const h of history) {
 				if(h.role === 'tool' && toolTimings.has(h.id)) Object.assign(h, toolTimings.get(h.id));